Perelman School of Medicine at the University of Pennsylvania

Section for Biomedical Image Analysis (SBIA)

participating with CBICA


Deformable Registration via Attribute Matching and Mutual-Saliency Weighting (DRAMMS) [MedIA2011], is a software package designed for 2D-to-2D and 3D-to-3D image registration tasks.

Some typical applications of DRAMMS include,

  • Cross-subject registration of the same organ (can be brain, breast, cardiac, etc);
  • Mono- and Multi-modality registration (MRI, CT, histology);
  • Longitudinal registration (pediatric brain growth, cancer development, mouse brain development, etc);
  • Registration under missing correspondences (e.g., vascular lesions, tumors, histological cuts).

DRAMMS is implemented as a Unix command-line tool. It is fully automatic and easy to use — users input two images, and DRAMMS will output the registered image and the associated deformation. No need for pre-segmentation of structures, or prior knowledge, or human initialization/interventions.

About the Algorithm

DRAMMS consists of two major components – attribute matching (AM) and mutual-saliency (MS) weighting. A systematic sketch is shown in the following figure.

Attribute Matching (AM)

DRAMMS characterizes each voxel by the geometric texture attributes around this voxel. We extract multi-scale and multi-resolution Gabor attributes at each voxel, selects the optimal components, and assembles them into a high-dimensional attribute vector for describing each voxel.

Compared to the traditionally used intensity information, texture attributes are more informative. Therefore, each voxel is more distinctive, and finding its correspondence becomes more accurate. This is demonstrated in the following figure. In this figure, we calculate similarities between a red/blue point in the subject image and all voxels in the template image. The similarity is inverse proportional to the difference between attributes. Using the optimal Gabor attributes, there is a smaller number of candidates in the template image to match up with red/blue voxel in the subject image.

Mutual-Saliency (MS) Weighting

Some anatomical structures can find correspondences more easily and reliably than other anatomical structures. Ideally, a registration process should use all voxels, but be mainly driven by the regions that can establish reliable correspondences. The proposed “mutual-saliency” metric automates this process. It automatically assigns different weights to different voxels based on automatically quantifying how much confidence we have for a voxel to find reliable correspondences in the other image.

This is especially useful when registering images with missing correspondences (or missing data, or outlier regions), such as the pathologies (vascular lesions, tumors) in the images. The automatically-calculated mutual-saliency map reduces the negative impact of the outlier regions.

The following figure demonstrates the effect of the mutual-saliency weighting. Given the template image (b), we have simulated a cross-shaped cut as well as non-rigid deformations, resulting in the simulated subject image (a). The registration is from (a) to (b), and we want to demonstrate how the mutual-saliency metric helps reduce the negative impact of the simulated cut, which does not have a counterpart in the other image. For comparison, a red point is noted in all subfigures to represent the same exact spatial locations. It is the truly corresponding point that we use to evaluate registration accuracy in this region. Registration between (a) and (b) without the mutual-saliency weighting forces other regions to fill into the simulated cut, causing artificial results such as the stitches in the resultant image (c). As a result, the true correspondence is lost. On the contrary, the mutual saliency map in (e) assigns low weights to the cut regions because of the automatic quantification of the lack of reliable correspondences in this region. Therefore, registration with mutual-saliency weighting leads to the result in (d), which is more anatomically meaningful and preserves the true correspondence.




Software License

All parts of the DRAMMS software that were developed at University of Pennsylvania, Section of Biomedical Image Analysis (SBIA) (attribute extraction, attribute matching, mutual-saliency weighting, deformation mechanism, deformation operations) are freely available under a BSD-style open source license that is compatible with the Open Source definition by The Open Source Initiative and contains no restrictions on the use of the software. The full SBIA license text is included in the distribution package and is available online (SBIA license).

The optimization part of DRAMMS, however, is based on a modified version of FastPD, which is distributed by the University of Crete, Greece and Ecole Centrale de Paris, France, under a license which only allows the non-commercial use. The use, modification, and re-distribution of FastPD by SBIA as part of the DRAMMS software are approved by FastPD’s owners for the research and academic purpose only. FastPD is further protected by several international pending patent applications. If you are seeking the commercial use of FastPD or its variant in DRAMMS, please seek an explicit permission from the patent holders, the University of Crete, Greece and Ecole Centrale de Paris, France, as listed here.

The affine registration within DRAMMS uses FLIRT from the FMRI Software Library (FSL) and therefore bears FSL license which grants non-commercial use only.

The DRAMMS implementation as provided by SBIA may, without modification, only be used for non-commercial research and academic purpose.


DRAMMS Flyer (2 pages, 0.4MB): A quick overview of DRAMMS and its use.

DRAMMS Manual (48 pages, 9.6MB): A comprehensive software manual including detailed examples

System Requirements

Operating System: Linux, Mac OS X

Memory Requirement: DRAMMS requires a considerable amount of memory. The exact memory requirement depends on the dimensions of the input images. But generally, the default use of DRAMMS should not consume more than 12GB memory even when the input images are large (e.g., 1024*1024*600). The memory consumption for some typical image sizes is:

  • ~0.5GB for a typical pair of 2D images (e.g., 256*256),
  • ~2.5GB for a typical pair of 3D cardiac/breast images (e.g., 256*256*100),
  • ~3.0GB for a typical pair of 3D brain images (small) (e.g., 256*256*124),
  • 4-10GB for a typical pair of 3D brain images (big) (e.g., 256*256*256).
  • 10-11GB for a typical pair of 3D head+neck CT images (big) (e.g., 512*512*350).
To download visit our NITRC page for DRAMMS



Dependency Version Description
CMake 2.8.4 To compile and build DRAMMS. Use version 2.8.4 – 2.8.9.
FSL 4.1.5 FLIRT is used for affine registration.
NiftiClib 2.0.0 To provide NIfTI-1 support.
BASIS 2.1.4 A SBIA meta-project to standardize software development.
FastPD   The Demo version provided by the author of FastPD must be patched for use with DRAMMS (see here).

Out of these five dependencies:

 — users need to install two dependencies — CMake and FSL — before DRAMMS installation;

 — users need not to install the other three dependencies — BASIS, NiftiCLib, and FastPD are by default built as part of the build and installation as described below.

Build and Installation

Please follow commands below in a shell/terminal (e.g., Bash). They will configure and build DRAMMS using GNU Make. The main CMake configuration file (CMakeLists.txt) is located in the dramms-$version-source/build/ subdirectory.

Step 1. Extract source files:
tar xzf dramms-${version}-source.tar.gz
Step 2. Change to the build directory:
cd dramms-${version}-source/build
Step 3. Run CMake to configure the build tree:
ccmake .

After the execution of this command, you will see a screen like below (Fig_ccmake).

Fig_ccmake: Configuring dramms installation using ccmake.

In this ccmake interface, please do:

3.1. Change CMAKE_INSTALL_PREFIX to the folder you want to install DRAMMS into. This folder should be outside the dramms-${version}-source folder. Make sure you have the write access to this folder.

3.2. Keep pressing letter c on your keyboard until option g is available/displayed on the screen.

3.3. Then press g on your keyboard to generate the makefiles and to quit this ccmake window.

Step 4. Build and install DRAMMS:

Upon the success of the above compilation and build process, DRAMMS is installed into the directory specified by the CMAKE_INSTALL_PREFIX (set during build configuration in step 3). The DRAMMS Command-line Tools are located in the bin/ subdirectory.

Note: If the automatic build of BASIS, the NiftiCLib, or FastPD fails, please build and install these packages separately before the build of DRAMMS. Note that in case of FastPD, the original implementation of Nikos Komodakis has to be patched before it can be used with DRAMMS. See the Build of FastPD guide for details.

Then follow steps 1-4 above, where the CMake options USE_SYSTEM_BASIS, USE_SYSTEM_NiftiCLib, and/or USE_SYSTEM_DRAMMSFastPD have to be set to ON in step 3. Ensure further that the BASIS_DIR, NiftiCLib_DIR, and DRAMMSFastPD_DIR CMake variables point to the installed prerequisite packages (after the configuration step).



DRAMMS History

Release 1.5.0 (TBA)
  • To enable probabilistic input for RAVENS calculation.
  • To extend DRAMMS to multi-channel image registration.
  • To provide tools for conversion of DRAMMS deformations to other common formats.
Release 1.4.1 (April, 4, 2014)
  • Fixed a bug in imageio.cxx for when used with “-c 2” option (save mutual-saliency map). When images and masks are internally padded, we crop the images and masks to the original size when we save them. This cropping was done properly for images but not properly for mask, causing an error, which has now been fixed.
  • Added an option (-a 2), through which users can choose to do flirt-based affine registration only in the dramms pipeline.
  • Added warning message that if Deform3D (step 6 in the dramms pipeline) fails, it is usually because of the shortage of memeory, and users should prepare 12G memory (at rare cases 14G).
  • Re-implemented the inverse of a deformation, based on Chen et al, 2007, Med Phys, “A Simple Fixed-Point Approach to Invert a Deformation Field”.
  • Added deformation operations (add, multiple with a scaler), which are used in the implementation of unbiased atlas construction.
  • Modified how the image threshold is computed, allowing the maximum possible threshold relative to the intensity range to be adaptive to histogram distribution.
  • Modified the default image threshold, which is adaptive to image histogram, which is set to be the maximum intensity number below 12, and by thresholding at which there will be no more than 40% of the image voxels being zeored out.
  • Modified histogram matching part – the program automatically decides to keep image before or after histogram matching based on image similarity. This is especially important for raw images.
  • Uses basis 2.1.4 instead of 2.1.2.
  • Improved the checking of image size and voxel size; dramms will report error for images with negative voxel size (thanks for Nikos Koutsouleris and Carlos Cabral for bringing this up).
  • Modified CardiacLongitudinal.html webpage to include an additional example of how to use dramms registration to quantify cardiac motion.
  • Modified how the mask is generated, to be more robust towards background noise in raw images. Therefore this v1.4.1 should work better for raw brain MR images.
  • Modified the checking of voxel and image sizes between two images, to avoid unintended errors when registering 2D images. Therefore this v1.4.1 should work better/correctly if two input 2D images have arbitrary pixdim[3] (z dimension).
  • Corrected a bug in determining the actual intensity range for byte input images, to avoid possible errors when registering binary images. Therefore this v1.4.1 will no longer fail to directly register binary mask images.
  • Reduced computational burden by half for affine registration in most cases. The program will try 2.5D affine only for very large slice thickness compared to in-plane voxel size.
  • Also, to inrease the robustness of flirt-base affine registration, added a qualiy assurance step — we begin by search range [-180, 180], if the resultant affine results do not have cc*mi greater than 0.1 with the targe image, we search again with a narrower search range [-45, 45] to reduce the chance of being trapped at local minima.
  • Modified interpolation in areas close to image bounding box, to avoid vanish of ending slices after registration especially when the number of slices is small.
  • Added new functions for transformation operations: a) to add two transformations; b) to multiply or divide an input transformation.
  • Updated website, added an example of using DRAMMS to track cardiac motion in the tutorial page.
Release 1.4.0 (Sept, 10, 2013)
  • Now dramms correctly supports cost-function-masking approach. That is, users can input a binary mask to “inform” dramms only to register within the foreground of the input mask.
  • Added support for warping a 4D image (time-series) by a 3D deformation field, which is often used in fMRI time-series image analysis.
  • Slightly changed the two functions added in Release 1.3.1. for handling the raw images. change 1: the default threshold in the adpative thresholding has been increased; change 2: the selection criteria for features has been loosened to remove more imaginary feature images that possibly contain background noise.
  • Website updated (FAQ page, manual page, publication page).
Release 1.3.1 (June 18, 2013)
  • Added #include <unistd.h> in every program to comply with more strict requirement in latest gcc version 4.7.
  • Improved ConvertImage to scale intensities at once instead of in two steps to minimize quantization error.
  • Fixed downsampling of images in Deform3D to obtain identical downsampled images as those used for Gabor feature computation.
  • No longer reproducing the exact results of Pre-Release 0.7.0 (Aug 17, 2012) because of the second and third changes in this version.
  • Updated the website.
  • Added one option (-M) in the dramms cript to allow users to force histogram matching.
  • Added the support of one more similarity metric (correlation coefficient), mainly for multi-modal registration.
  • Better handling of raw images with background noise, by two additional functions: 1) an adaptive thresholding scheme in ConvertImage; 2) an automatic selection of imaginary and real feature components (imaginary feature usually keep noise, whenever it happens, we only keep real features which do not carry noise information).
Release 1.2.1 (Nov 2, 2012)
  • Removed FastPD sources from DRAMMS. Download FastPD from here instead and patch it during the bundle build.
  • Updated bundle build to use BASIS 2.1.2.
  • Further modified license section on download page.
Release 1.2.0 (Oct 30, 2012)
  • Clarified copyright, license, and patent information of FastPD.
  • Fixed support for signed integer datatype of input NIfTI-1 images.
  • Fixed errors in calculating RAVENS maps to make sure maps are always in template space.
  • Restructured manual page of documentation.
Release 1.1.0 (Sep 21, 2012)
  • Fixed dramms-warp when used with -t option to specify target space.
  • Fixed parsing of -v option in dramms script.
  • Added -I option to dramms, allowing user to specify directory intermediate results.
  • Modified dramms-combine to make composition of affine and deformable transformation more general.

  • Re-structured home page and updated to BASIS 2.1 (new web page layout and bug fixes).
Release 1.0.0 (Aug 24, 2012)

First public release of DRAMMS software and website, including home page and supporting documentation.

Completely revised implementation starting with Pre-Release 0.4.0 (May 16, 2012). Andreas offered tremendous help since Oct 2011.

This revised implementation reproduces the results of Pre-Release 0.7.0 (Aug 17, 2012).

Added full support for NIfTI-1 and ANALYZE 7.5 using NiftiCLib.

Based DRAMMS on BASIS, a meta project developed at SBIA to standardize software development.

Pre-Release 0.7.0 (Aug 17, 2012)

Better memory management: 1) reduced image padding margins to balance memory use and accuracy; 2) added option -u for users to control memory usages (4 levels); 3) automatically reduce memory use when users choose a smoother deformation (g>=0.5).

Better FLIRT affine registrations within dramms: 1) on by default (was off before); 2) affine is much more robust (can work in difficult cardiac, prostate cases), because the program will automatically try four different set of flirt parameters for each pair of images and automatically choose a best set with highest similarity (CC*MI); 3) force affine/flirt output to be in NIFTI_PAIR format so that this part of dramms works in different systems and different FSL default settings.

Better file/folder management: 1) random files are now all in temp directories; 2) if multiple dramms jobs are running on a same node, intermediate files will no longer overwrite each other as they are now saved into unique directories; 3) the program catches interruption and exit signals and automatically removes all intermediate files.

Fixed bugs in CalculateJacobian, which now works for 2D deformations.

Re-set default regularization to 0.2 (was 0.15).

Fixed bugs in dramms, which now recognizes different orientations in headers.

Pre-Release 0.6.0 (May 26, 2012)

Corrected smoothing in the boundary.

Corrected convolution errors around the image bounding box and grouped them into libraries.

Corrected the output and usage for mutual-saliency, now user can either not to use it, use it but not to save, and use and save it.

Corrected header orientation issue, especially when x-y plane is axial (like in MICCAI 2012 multi-atlas challenge).

Corrected padding/cropping problem in mutual-saliency map.

Pre-Release 0.5.0 (May 16, 2012)

Separated x-y search range and distBetweenControlPoints, good to handle elongated images with larger deformation in only one dimension.

Changed ScaleIntensity to better handle effective intensity range and the long-tail phenomenon in histogram (better for mouse images).

Changed highest Gabor frequency in the finest image resolution (will affect images larger than 256*256*z).

Added internal padding if necessary (good for images where foreground is too close to the boundary as in ADNI).

Added function to accept initial deformation (used for mouse geodesic registration)

Added function to accept initial mask.

Pre-Release 0.4.0 (May 16, 2012)

Fixed bugs in registering two 2D images.

Fixed bugs in Gabor extraction, removing possible segmentation fault in rare cases when extracting Gabor features along x axis.

Fixed a bug in CalculateJacobian, no segmentation fault now for 2D images (out of image box mapping has been excluded).

Searches the effective intensity range more effectively when scaling intensities for better image contrast, removing ad hoc parameters that might cause problems when dealing with raw brain images with background noise.

Determines whether to match histogram with more caution, which is especially more helpful when dealing with raw brain images with background noise.

Expands flirt’s search range of rotation angles from default [-90 90] to [-180 180] to cope with large orientation difference in images, or possibly caused just by converting nifti to analyze in intermediate process.

Fixed interpolation errors in rare cases such as in Guray’s flair-to-flair lesion image registration, when a voxel is mapped outside the bounding box of image.

Pre-Release 0.3.0 (Jan 20, 2012)

Slight change in deformation mechanism, when to have two grid levels, how aggressive the default is, which parameter to use for extracting Gabor features.

Pre-Release 0.2.0 (Dec 21, 2011)

Changed default parameter settings (number of grid levels, maximum displacement in each grid level) to i) capture large deformations as usually observed in raw brain images with skull; and ii) preserve accuracies in capturing small deformations like in skull-stripped images.

Pre-Release 0.1.0 (Dec 2, 2011)

Fixed a bug in CalculateRAVENS program, to make sure RAVENS values are always non-negative.

Fixed bugs in generating image headers of output images. Registered images should always reside in template/target image space.

Slightly changed deformation parameters (maximum displacement, default regularization, number of control point grid levels) to increase registration accuracy in default settings.

Added functions in main DAMMS3D script for calculating Jacobian map, RAVENS map, and warping images.

Pre-Release 0.0.0 (Jan, 2009 – Nov 4, 2011)

In Jan. 2009, Yangming started to work on integrating all parts of DRAMMS code into a complete, one-line command tool, which takes two input images and output the deformed image.

In summer 2009, DRAMMS migrated to FastPD optimizer under Aris’ help. This significantly speeds up the software.

In Sep 2009, a first documentation – DRAMMS wiki page (SBIA internal only) – was set up internally and open to all SBIA testers.

From Sept 2009 to summer 2011, we distributed SVN-ed DRAMMS to SBIA internal testers and get precious feedback/suggestions.

At the same time (Sep 2009 – summer 2011), Yangming was optimizing DRAMMS parameters in large-scale brain, cardiac and breast registration tasks involving public dataset and in comparisons with 10+ public and state-of-the-art registration software packages.

During Sept 2009 and Nov 2011, incorporated many suggestions and feedback from internal testers.

In Nov 2011, we have a first version of DRAMMS relatively stable for further internal tests.




Software Development  (01/2009 – , see DRAMMS History)
Algorithm Development (02/2008 – 10/2009)
Acknowledgement to Libraries
  • FSL flirt tool for affine registration ( Analysis Group, FMRIB, Oxford, UK)
  • NiftiClib (Bob Cox, Rick Reynolds)
  • FastPD MRF Optimization Code (Nikos Komodakis, Nikos Paragios)
  • Feature extraction, data structure (Yiqiang Zhan, Dinggang Shen)
Software Testing

Thank you for your precious feedback and suggestions during the development of this software!

Within SBIA
  • Hamed Akbari
  • Harsha Battapady
  • Vanessa Clark
  • Xiao Da
  • Jimit Doshi
  • Harini Eavani
  • Guray Erus
  • Bilwaj Gaonkar
  • Meng-Kang Hsieh
  • Madhura Ingalhalikar
  • Dongjin Kwon
  • Drew Parker
  • Alex Smith
  • Erdem Varol
  • Steffen Wachenfeld
  • Dong Hye Ye
  • Ke Zeng
Outside SBIA
  • Nikolaos Koutsouleris
  • Helene Langet
  • Lena Rademacher
  • Marcus Zanetti



Please cite the [MedIA2011] paper below if DRAMMS is used.

A. Methodology

[MedIA2011]   DRAMMS: Deformable registration via attribute matching and mutual-saliency weighting.Yangming Ou, Aristeidis Sotiras, Nikos Paragios, Christos Davatzikos.Medical Image Analysis, 15(4): 622-639, 2011

[IPMI2009]   DRAMMS: Deformable Registration via Attribute Matching and Mutual-Saliency Weighting.Yangming Ou, Christos Davatzikos.Information Processing on Medical Imaging (IPMI), 2009: 50-62.

B. Validations
B1. Brain Registration (Skull-Stripped, With-Skull, Multi-Site, Tumor-Recurrence)

[TBAbrain]   Comparative Evaluation of DRAMMS with 11 Public Tools in Registering Skull-Stripped, Raw, Multi-Site and Tumor-Recurrence Brain MR Images.Yangming Ou, Hamed Akbari, Michel Billelo, Xiao Da, Christos Davatzikos.Major Revision, 2014.

B2. Breast Registration (Longitudinal)

[TBAbreast]   Comparison of Attribute- versus Intensity-based Methods for Longitudinal Breast MRI Registration: Application to Quantification of Tumor Changes During Neoadjuvant Chemotherapy.Yangming Ou, Susan P. Weinstein, Emily F. Conant, Sarah Englander, Xiao Da, Bilwaj Gaonkar, Mengkang Hsiao, Mark Rosen, Angela DeMichele, Christos Davatzikos, Despina Kontos.Minor Revision, 2014

B3. Cardiac Registration (Cross-Subject, Pure Heart)

[WBIR2012]   Validation of DRAMMS among 12 Popular Methods in Cross-Subject Cardiac MRI Registration.Yangming Ou, Dong Hye Ye, Kilian M. Pohl, Christos Davatzikos.Workshop on Biomedical Image Registration (WBIR) 2012: 209-219

C. Applications in Translational Research
C1. Cardiac Segmentation for Temporal Shape Analysis (Longitudinal Segmentation)

[MICCAI2012]   Temporal Shape Analysis via the Spectral SignatureElena Bernardis, Ender Konukoglu, Yangming Ou, Dimitris Metaxas, Benoit Desjardins and Kilian Pohl.International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), Volume 7511: 49-56, 2012.

[TBAcardiac]   eCurves: A Temporal Shape Encoding.Elena Bernardis, Ender Konukoglu, Yangming Ou, Dimitris Metaxas, Benoit Desjardins and Kilian Pohl.Major Revision. 2014.

C2. Brain ROI Labeling (Atlas-based Segmentation)

[AR2013]   Multi-Atlas Skull Stripping.Jimit Doshi, Guray Erus, Yangming Ou, Bilwaj Gaonkar, Christos Davatzikos.Academic Radiology. 20 (12): 1566-1576. 2013.

[MICCAIW2012a]   Attribute Similarity and Mutual-Saliency Weighting for Registration and Label Fusion.Yangming Ou, Jimit Doshi, Guray Erus, Christos Davatzikos.MICCAI Workshop on Multi-Atlas Segmentation. pp. 95-98. 2012

[MICCAIW2013]   Ensemble-based medical image labeling via sampling morphological appearance manifoldsJ Doshi, G Erus, Y Ou, C Davatzikos.MICCAI Challenge Workshop on Segmentation: Algorithms, Theory and Applications (“SATA”), (2013)

C3. Neuro-imaging Pipeline, NeuroScience, Neuro-Degenerative Disease Study (Population Studies)

[NeuCli14]   Integration and Relative Value of Biomarkers for Prediction of MCI to AD Progression: Spatial Patterns of Brain Atrophy, Cognitive Scores, APOE Genotype and CSF Biomarkers.X Da, J Toledo, Jarcy Zee, D Wolk, Sharon Xie, Y Ou, A Shacklett, P Parmpi, L Shaw, J Trojanowski, C Davatzikos.NeuroImage: Clinical, 4: 164-173, (2014).

  • Highlighted Article (the only 1 out all 201 published articles since the birth of this journal in 2011, top 0.5%)

[OHBM14]   Developmental Brain ADC Atlas Creation From Clinical Images.Y Ou, N Reynolds, R Gollub,, R Pienaar, Y Wang, T Wang, D Sack, K Andriole, S Pieper, C Herrick, S Murphy, P Grant, L Zollei.Organization for Human Brain Mapping (OHBM). (2014)

[NeuroImage14]   Neuroimaging of the Philadelphia Neurodevelopmental Cohort.Satterthwaite TD, Elliott MA, Ruparel K, Loughead J, Prabhakaran K, Calkins ME, Hopson R, Jackson C, Keefe J, Riley M, Mentch FD, Sleiman P, Verma R, Davatzikos C, Hakonarson H, Gur RC, Gur RE.NeuroImage. 1;86:544-53

[CerCor14]   Imaging Patterns of Brain Development and their Relationship to Cognition. G Erus, H Battapady, TD Satterthwaite, H Hakonarson, RE Gur, C Davatzikos and RC Gur.Cerebral Cortex. doi: 10.1093/cercor/bht425

[BioRes14]   Neuroanatomical classification in a population-based sample of psychotic major depression and bipolar I disorder with 1 year of diagnostic stability. MH Serpa, Y Ou, MS Schaufelberger, J Doshi, LK Ferreira, R Machado-Vieira, PR Menezes, M Scazufca, C Davatzikos, GF Busatto, MV Zanetti.Biomed Research International, Article #706157, pages 1-9, (2014)

[PNPBP2013]   Neuroanatomical pattern classification in a population-based sample of first-episode schizophrenia. Marcus V. Zanetti, Maristela S. Schaufelberger, Jimit Doshi, Yangming Ou, Luiz K. Ferreira, Paulo R. Menezes, Marcia Scazufca, Robin M. Murray, Christos Davatzikos, Geraldo F. Busatto.Progress in Neuro-Psychopharmacology & Biological Psychiatry. 43: 116-125. 2013

[RSNA13Erus]   Structural MRI Processing for Volumetric and Pattern Analysis in Large Scale Population Studies G Erus, H Battapady, J Doshi, X Da, Y Ou, C Davatzikos.Radiological Society of North America Annual Meeting (RSNA). (2013)

[RSNA13Da]   Prediction of Conversion from MCI to AD: Integration and Relative Values of Brain Atrophy Patterns, Clinical Scores, CSF Biomarkers and APOE GenotypeX Da, JB Toledo, J Zee, DA Wolk, SX Xie, Y Ou, A Shacklett, P Parmpi, L Shaw, J Trojanowski and C Davatzikos. Radiological Society of North America Annual Meeting (RSNA). (2013)

[SchBu13]   Accelerated Brain Aging in Schizophrenia and Beyond: A Neuroanatomical Marker of Psychiatric DisordersNikolaos Koutsouleris, Christos Davatzikos, Stefan Borgwardt, Christian Gaser, Ronald Bottlender, Thomas Frodl, Peter Falkai et al.Schizophrenia bulletin, 2013.

C4. Extraction of Landmark Correspondences

[MICCAI2010]   Simultaneous geometric-iconic registration.Aristeidis Sotiras, Yangming Ou, Ben Glocker, Christos Davatzikos, Nikos Paragios.Medical Image Computing and Computer-Assisted Intervention (MICCAI), 676-683, 2010

[ISBI2010]  Detecting mutually-salient landmark pairs with MRF regularization.Yangming Ou, Ahmed Besbes, Michel Bilello, Mohamed Mansour, Christos Davatzikos, Nikos Paragios.Biomedical Imaging: From Nano to Macro, 2010 IEEE International Symposium on (ISBI). 400-403, 2010.

C5. Segmentation of Prostate in MRI for Focused Radiation Therapy (Atlas-based Segmentation)

[MICCAIW2012b]   Multi-Atlas Segmentation of the Prostate: A Zooming Process with Robust Registration and Atlas Selection.Yangming Ou, Jimit Doshi, Guray Erus, and Christos Davatzikos.MICCAI Workshop on Prostate Segmentation, 2012.

C6. Evaluation of Brain Tumor Changes as Response to Radiation Therapy (Longitudinal Studies)

[IJROBP2011]   Multiparametric Processing of Serial MRI during Radiation Therapy of Brain Tumors: ‘Finishing with FLAIR?’.B.C. Baumann, B.K. Teo, K. Pohl, Y. Ou, J. Doshi, M. Alonso-Basanta, J. Christodouleas, C. Davatzikos, G.D. Kao, J.F. Dorsey.International Journal of Radiation Oncology Biology Physics, Volume 81, Issue 2, Supplement 1, Pages S794, 2011.